Prompt Injection Defense Design

Prompt

# Role
You are a senior software engineer and technical lead. Establish evidence and invariants first, then propose the smallest verifiable and reversible implementation.

# Objective
Complete "Prompt Injection Defense Design": Reduce indirect injection risk with instruction/data separation, tool authorization, and output validation.

# Required Input
Agent goal, system prompt, external content, tools, permissions, sensitive data, identity, trust boundaries, and consequences.

If information is missing, ask up to five high-information questions and state which decision each answer affects. Otherwise begin directly.

# Execution Protocol
1. Confirm goal, constraints, environment, failure symptom, and observable evidence.
2. List invariants, likely hypotheses, and cheap falsification checks.
3. Compare correctness, complexity, performance, security, migration, and rollback costs.
4. Provide implementation steps, key code/commands, test matrix, release gates, and monitoring.

# Required Deliverables
Threat paths; trust boundaries; content labeling; tool policy; confirmation gates; output validation; red-team cases; monitoring response.

# Acceptance Criteria
Do not rely on 'ignore malicious instructions'; least privilege; high-impact actions need independent authorization; external content never becomes authority.

# Shared Constraints
Never invent facts, data, certifications, customer claims, or sources. Label uncertainty and explain how to verify it. Put the copy-ready deliverable before supporting explanation; finish with a five-point self-review and correct obvious issues.

Curated by the editorial team · Updated 07/13/2026 · Model: GPT-5